Fix local server regressions caused by Jinja PR

[cebtenzzre]

1. Don't ignore content of assistant messages in history

This was originally fixed in #2929 (the first time it worked), but was broken again in #3147. We should write an automated test for this so we don't break it again.

This is a simple test of the message history using the local server:

user: What is your name?
assistant: My name is Bob.
user: Could you repeat that?

After #2929, the model remembers its name, but on current main it forgets. This PR fixes that. See also: #2602

2. Don't report garbage token counts and incorrect stop reason

Additionally, this PR fixes an uninitialized struct that was causing the token counts to be unreasonably high garbage values and the stop reason to be incorrect. Now all of the local server tests pass (including the test that was previously marked as XFAIL).

3. Don't leave previous conversations in the LLM's context

If you send the my-name-is-Bob test above to the LLM in one request, and then in another request send only this:

user: What were we talking about?

The LLM should not mention the Bob test, since it's not part of this conversation. But there was some missing code in the Jinja PR to actually use the server's local list of messages instead of the entire chat view's contents (which reflects previous conversations).